Capstone: Regional Referral Hospital Stroke Risk Pipeline & Explainable AI (XAI)

The capstone joins every module into one service for a high-volume Ugandan Regional Referral Hospital:

1. Data — the stroke dataset with leakage-free preprocessing in an sklearn Pipeline (module 1).
2. Artifact — the whole pipeline saved with joblib, versioned and checksummed (module 2).
3. API — FastAPI + Pydantic /predict, model loaded once per worker (module 3).
4. Infrastructure — CPU instances in a nearby region, Spot only for retraining (module 4).
5. Container — multi-stage, non-root Docker image (module 5).
6. Release — canary with automatic rollback (module 6).
7. Performance — ONNX/INT8 if latency demands it (module 7).
8. Monitoring — PSI drift detection and delayed-label evaluation (module 8).

What the capstone adds: choosing the decision threshold for clinical use, and explaining each prediction to the physician.

A classifier outputs a probability; a threshold turns it into an action (refer to the chronic-care clinic or not). The default 0.5 is arbitrary — and terrible when only ~5% of patients have strokes, because most stroke patients get probabilities well below 0.5.

Lowering the threshold trades precision for recall:

The right threshold balances the cost of a missed stroke against the clinic's capacity to see referred patients. Choose it on a validation set (never the test set), document why, and revisit it when prevalence or capacity changes. Per the Ministry of Health integration guidelines (June 2025), high-risk patients at triage are enrolled in the chronic care clinic — so the threshold directly sets the clinic's workload.

"This patient's stroke risk is 42%" is not enough for a referral physician. They need to know why — to check the model against their own judgement, catch data errors (an age typed as 670), explain the decision to the patient, and decide what to treat.

Explainability also supports accountability and fairness: you can audit whether the model leans on features it shouldn't, and justify decisions to patients and regulators. In clinical settings an unexplained score is often simply ignored.

SHAP (SHapley Additive exPlanations) assigns each feature a contribution to this prediction, based on Shapley values from cooperative game theory. The contributions are additive:

prediction = base value + Σ feature contributions

where the base value is the model's average output (≈ the hospital's base rate). For a patient: "base risk 5% → this patient 42%: +18 points from age 68, +12 from hypertension, +7 from glucose 220 mg/dL…".

Key properties for exams:

• Local — explains one prediction; averaging |SHAP| across patients gives global importance.
• Additive and consistent — contributions sum exactly to the prediction (in the model's output space, e.g. log-odds for logistic/XGBoost models).
• TreeSHAP computes exact values efficiently for tree models like XGBoost.
• SHAP explains the model, not causality: "+12 from hypertension" means the model's prediction rose, not that hypertension causes 12 points of risk for this person.